Lets take the GPU competition argument who's better to one of [Police Academia quote]The many, many... many many..." other GPU dedicated threads already existing.

Thanks.

And no, there is absolutely no GPU client or science app planned to run for NRW on WCG. Read the last news update(s) on lack of funding!

Again, not true. Collatz and distributed.net (and dnetc and moo! wrapper, that just crunch the dnet work on boinc) are single precision and ATI is much faster than Nvidia.

About Folding, their ATI client don't fully use the card, they stated that their current proteins don't use all the shaders of it.

Any new updates?

I guess it's time for me to jump in at this point with an update on several issues for NRW. It is not dead.

Science:

The original intention was just to put up the best structures for each sequence that was there and give a best guess at the function based on the similarity of the best structures to the existing protein structures where we know the function. We decided that this was not going to be useful in itself without some quantification of how likely the predictions were correct. Problem, is that there are almost no experimentally derived structures with which to make a comparison set where we *know* the answer.

So we did the next best thing, take the sequences which had a decent degree of sequence similarity to known structures and make homology models - basically superpose the existing structure onto the new sequence. With good sequence matches this is a very accurate technique and provides with a set of structures to compare the NRW predictions and gauge the accuracy. So we did this with the 2000 best matching sequences (out of the 62000) and found that for almost 38% of the sequences, NRW had gotten a similar result in terms getting the right family of folds, at least for a sub-domain of 40 residues or more.

This is a great result - but we need to be able to both identify the correct substructure and to identify when it was correctly predicted and when it was not, or at least come up with a confidence value. This is what we are doing now.

GPUs

We have continued with the GPU work. I have ported a lot of our ATI stuff from Brook+ to OpenCL which is a much nicer language once you find the subset that is actually supported and figure out what the GPU is actually doing under the hood. We are actively using GPUs in the analyses. The ProtinfoAB client has long been ported to GPUs and thoroughly tested as a Brook+ app and it is on the short list of things to port to OpenCL. I shouldn't say port actually since it is quite different from the original client with many improvements. Thats the main reason it's taking time - not so much the actual coding, but improving the algorithm which takes priority over writing it in OpenCL. Haven't yet played with the NVIDIA flavor of OpenCL but we have some new NVIDIA GPUs for this purpose.

Funding

The funding environment right now is very poor. Even the best applications are getting rejected now which is a real pity. For us, we need to publish the results of NRW before we can ask for money to continue the work with 1KP. This is reasonable. At the same time, we don't want to make the data as useful as possible before putting it out. There is quite a bit of skepticism about how well this approach will work. This again is fair, but again, I think that explains the motivation in making sure that the function predictions are the best that we can make them before putting them on the website.

I hope this answers some of your questions and I will be updating the website soon.

Hong

Thanks for the update